TencentDB Agent Memory
A tiernew this weekAn open-source, fully local hierarchical memory system for AI agents that cuts token bloat and survives long-horizon sessions — no external API required.
Kai's verdict
One of the more technically serious open-source agent memory solutions out there — the layered semantic pyramid is a genuine architectural improvement over flat-vector approaches, and the benchmark numbers are credible. Still early and framework-coupled, but worth watching closely if you're building production agents. (Verdict pending Phi's full review.)
Strengths
- 4-tier semantic memory pyramid (L0 Conversation → L1 Atom → L2 Scenario → L3 Persona) beats flat vector stores at recall quality
- 61.38% token reduction with full traceability preserved — not lossy summarization
- Zero external API dependencies; runs on local SQLite + sqlite-vec by default
- Hybrid BM25 + vector + RRF retrieval gives more reliable recall than pure vector search
- Strong benchmark gains: SWE-bench 58.4% → 64.2%, PersonaMem accuracy 48% → 76% in long-horizon sessions
Weaknesses
- Tightly coupled to OpenClaw and Hermes agent frameworks; integrating with arbitrary agents requires extra work
- Self-hosted setup (Node.js 22+, Docker, SQLite) adds DevOps friction vs. a managed memory service
- Early-stage project (v0.3.x) — API and config surface is still shifting
Best for
Agent engineers building long-horizon coding or task agents who are hitting context window limits and need structured, debuggable memory without paying for a managed vector DB.
Pricing
Free (MIT open source)
Fully free and self-hosted; optional paid Tencent Cloud Vector Database (TCVDB) backend for scale